psyche-network-rules-reference

# Psyche network-rules reference `.rules` is a heartbeat signal, not a Psyche action or generic instruction API. The agent heartbeat (`_check_rules_file`) consumes it, unlike ordinary durable edits, which require file→`context.rebuild`. ## Write and consume Write the complete approved UTF-8 body atomically to the explicitly authorized agent workdir root (POSIX example; use the active shell’s equivalent elsewhere): ```sh target='/absolute/path/to/authorized-agent' body='/absolute/path/to/approved-rules.txt' tmp=$(mktemp "$target/.rules.XXXXXX") && cat "$body" > "$tmp" && mv "$tmp" "$target/.rules" ``` A failed command requires inspection of that exact temporary file; do not blindly replay or announce success. Confirm each target separately—there is no descendant broadcast. Do not substitute `system/rules.md` for this live signal. The next runnable heartbeat reads and unlinks `.rules` before deciding whether to apply it. Read/unlink failure leaves it unconsumed and stops processing, so disappearance alone is not proof of success. A nonempty body completely replaces `system/rules.md` and the protected `rules` section; whitespace-only content is a consumed no-op. Identical content is consumed without rewrite or flush. Changed content persists and flushes, logging `rules_loaded`; a canonical-file write failure aborts before prompt mutation and logs `rules_write_error`. Boot and every full rebuild/refresh/molt reread `system/rules.md` independently; a missing or empty canonical file removes the section. Identical comparison ignores leading/trailing whitespace. A pending `.rules` file is not yet applied, and an empty signal cannot clear existing rules. Verify the canonical file on disk and the effective protected section only after that agent's heartbeat processed the signal (or after later reconstruction). ## Boundary Writing another reachable agent's `.rules` file is an ordinary filesystem write, not a dedicated avatar authorization path. Reachability is not permission: only actual human scope and the target's real accessibility authorize the write. Keep this distinction visible when operating across agents.